arXiv:2608.19206v1 Announce Type: new Abstract: Contemporary Large Language Models (LLMs) are increasingly aligned to suppress hallucinations, prioritizing factual retrieval over combinatorial creativity. While crucial for mitigating misinformation, this alignment may also restrict speculative Research and Development (R&D) by encouraging what this work operationally treats as semantic overfitting and diversity collapse. In this paper, we propose a Rust-based multi-agent orchestration that uses the contrast between narrative daydreaming and executive control as a functional analogy, not as a neurocognitive claim. The system instigates an Epistemological Friction loop between a high-entropy generating agent and a web-grounded evaluating agent, mediated by a low-entropy semantic bottleneck intended to reduce noise and repetition. Initial experiments generated diverse, viability-rated hypotheses across physical and social-science domains. We additionally report an exploratory paired basel...